On Learning 3D Eye Vessel Reconstruction from Procedurally Generated Data

Human vision system lesions are miserably widespread. World Health Organization reports tens millions of diagnoses for human retina pathology. The following article addresses problem of retina image processing assisted with deep learning. We focus on advanced problem statements and we believe the third dimension is the key for these problem solutions. Neural networks used in this article fully trained with synthetic datasets as the genuine 3D retina data is currently unavailable. Datasets created with procedural retina vessel generator which is also developed for the purposes of present study. We show that neural networks can handle 3D reconstruction task applied to retina vessels with production-ready model fidelity level. We compare achieved results with another techniques for 3D reconstruction. We also initiate a discussion about properties and distributions of synthetic data compared to true human retina so the question of transferring results to patient health check application is also covered. Present study is the first step of synthetic data based approach to human eye vessel third dimensional reconstruction. Achieved results are to be used in future work which will be focused on using genuine retina images taken during ophthalmoscopy.